Predicting transformer failures before the lights go out

The problem

A utility operated tens of thousands of distribution transformers, most well past their nameplate life. Maintenance was a mix of fixed schedules and failure response. Fixed schedules replaced healthy assets and missed sick ones; failure response meant emergency crews, overtime, and customers in the dark. The asset data existed — SCADA telemetry, oil test results, load history, decades of work orders — but it lived in four systems that had never been joined.

The hypothesis

Failure is rarely sudden. It is preceded by months of thermal, load, and dissolved-gas signatures. If those signals were unified per asset, a model could rank the fleet by failure risk far enough ahead for the work to be planned rather than scrambled.

The build

• Asset-centric data foundation — telemetry, oil analysis, work orders, and weather joined to a single asset identity with full lineage. This was the majority of the work and the reason everything downstream was possible.
• Survival and gradient-boosted models — predicting probability of failure inside a rolling horizon, with time-to-event framing rather than a simple binary label.
• Reason codes on every score — each prediction carries its top contributing factors, so a planner sees rising top-oil temperature and gas trend, not an unexplained number.
• Integration into the work-management system — risk scores flow into the existing planning queue. No new tool for crews to learn; the ranking simply got better.

Rollout

Six months of backtesting against actual failures, then a shadow period where predictions were recorded but planning was unchanged. Only after the shadow period matched backtest performance did the scores start driving work. Crews were skeptical, reasonably — the reason codes were what convinced them, because they matched what experienced technicians already looked for.

Results

The utility gained months of median warning on at-risk assets, unplanned outages fell, and maintenance spend shifted from emergency to planned work at materially lower cost per intervention. The reliability metrics regulators care about moved in the right direction.

What we'd tell you

• The data foundation is the project. Modeling is the short part.
• Validate out-of-time. Random splits on time-series data flatter you and then fail you.
• Reason codes convert skeptical experts into users. A bare score does not.
• Deliver into the tool crews already use. A separate dashboard is a separate thing to ignore.